Compose Skill (Compose Kit)

A Compose skill for AI coding agents: seven agent skills that make Claude Code, Codex, Cursor, Copilot and Gemini write better Jetpack Compose and Compose Multiplatform code. One consistent house style, measured on seven models in real agent CLIs.

Works with Claude Code · Codex · Cursor · OpenCode · Copilot · Gemini CLI · Antigravity

Website: compose.avinya.dev · Install guides per agent · Results

What we achieved

• Held-out v5 (120 agentic runs): the kit lifted Gemini 3.8 Flash by +25.6 points and Sonnet 5.5 by +18.6, and the test found a real flaw: on a "match our conventions" task, models restructured too much with the kit (results).
• v6.0.0-preview.1 (2026-10-01): seven skills, one entry tree, four install channels.
• Website live (2026-10-02): compose.avinya.dev, with install guides per agent and a build check that every number on it appears in the evidence files.
• v6 A/B (90 agentic runs): v6.0 vs the v6.1 fixes vs no kit. The "no new UI controls, screens or features" item on conform tasks went from 2 of 10 cells (v6.0) to 8 of 10 (v6.1), and every pre-registered rule held (results).
• v6.1.0 stable (2026-10-05): no model scored below its own no-kit result, so it is released as stable. v6.1.1 (2026-10-11), the GitHub "Latest" release, keeps that skill guidance and fixes a guard-script security issue.
• Cheap models (36 add-on runs): DeepSeek V4.1 Flash went from 82.1% to 97.4% of items with v6.1, and DeepSeek V4 Pro from 74.4% to 87.2% (results).

Why

AI agents write Compose that compiles and then drifts: state ends up in the wrong place, every screen gets its own pattern, error handling changes from file to file, and reviews flag perfectly fine code as blocking. Compose Kit gives the agent one house style plus the reasoning behind it, and only loads the parts the current task needs.

• One consistent style: MVI on a shared BaseViewModel contract, Koin annotations, Navigation 3, feature-owned data/domain/presentation layers, and typed error handling.
• Progressive loading: a ~1.3k-token entry tree picks the task path and the exact reference files. The full kit is about 127k tokens; in the v5 test, 35 of 40 kit runs loaded under 20k.
• Respects your project: in a coherent existing codebase (Hilt, MVVM, Navigation 2, your own base class), the kit follows your pattern and proposes migration separately.
• Checked, not trusted: 11 shell guard scripts catch drift in CI, backed by 90 fixture tests.

What's inside